Hrv state evaluation method and device based on ppg data

Abstract

This disclosure provides a method and apparatus for HRV status assessment based on PPG data, comprising: acquiring a multi-mode physiological signal time series collected by a wearable device from a target object within a preset monitoring period, including ECG signal and PPG signal time series; performing successive heartbeat waveform structure segmentation processing on the ECG signal time series to obtain a heartbeat waveform unit sequence; and performing successive pulse waveform structure segmentation processing on the PPG signal time series to obtain a pulse waveform unit sequence; constructing a heartbeat waveform morphological feature evolution trajectory and a pulse waveform morphological feature evolution trajectory based on the heartbeat waveform unit sequence and the pulse waveform unit sequence; performing synchronous correlation and alignment processing on the heartbeat waveform and pulse waveform morphological feature evolution trajectory on the time axis; and training an HRV assessment network model based on this; and inputting the real-time acquired pulse waveform morphological feature evolution trajectory into the HRV assessment network model to output the corresponding HRV data.

Description

Technical Field

[0001] This disclosure relates to the field of health monitoring technology, and in particular to a method and apparatus for assessing HRV status based on PPG data. Background Technology

[0002] In the field of health monitoring technology, heart rate variability (HRV), as an important indicator reflecting autonomic nervous activity, is traditionally assessed using methods that heavily rely on electrocardiogram (ECG) signals. However, in wearable device applications, the high power consumption of circuits involved in ECG signal acquisition, amplification, and analog-to-digital conversion limits the battery life of wearable devices, making it difficult to meet users' daily monitoring needs for all-day, long-term use. Furthermore, while photoplethysmography (PPG) is widely used in wearable devices for HRV analysis due to its low power consumption and ease of integration, the waveform stability of PPG signals is poor because they are easily affected by factors such as ambient light, skin pigmentation, and motion. This results in HRV data acquired based on PPG signals often lacking accuracy and failing to accurately reflect the true state of cardiac autonomic nervous regulation. Moreover, traditional methods often analyze ECG or PPG signals in isolation, failing to fully explore the intrinsic correlation and synergistic changes in their waveform morphology evolution, leading to HRV assessment results that are unsatisfactory in terms of reliability and stability for clinical or health management needs. Summary of the Invention

[0003] In a first aspect, embodiments of this disclosure provide a method for HRV status assessment based on PPG data, the method comprising: Acquire the time series of multimodal physiological signals of the target object collected by wearable devices within a preset monitoring period; The ECG signal time series in the multi-mode physiological signal time series is processed by successive heartbeat waveform structure segmentation to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units; and the PPG signal time series in the multi-mode physiological signal time series is processed by successive pulse waveform structure segmentation to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units. The evolution trajectory of heartbeat waveform morphology features is constructed based on the QRS wave characteristics of each heartbeat waveform unit in the heartbeat waveform unit sequence; and the evolution trajectory of pulse waveform morphology features is constructed based on the slope characteristics of the ascending branch, the steepness characteristics of the descending branch, and the relative height characteristics of the diphthong notch of each pulse waveform unit in the pulse waveform unit sequence. The evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features are synchronously correlated and aligned on the time axis; and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous correlation and alignment are calculated. The HRV evaluation network model was trained using the evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features after synchronous correlation and alignment processing, as well as the HRV data corresponding to the evolution trajectories of heartbeat waveform morphology features after synchronous correlation and alignment processing. The time series of PPG signals collected by wearable devices for the target object are acquired in real time. The evolution trajectory of pulse waveform morphology features is constructed in real time based on the acquired PPG signal time series. The real-time constructed evolution trajectory of pulse waveform morphology features is input into the trained HRV evaluation network model, and the corresponding HRV data is output.

[0004] In some possible implementations of the first aspect, the ECG signal time series is preprocessed by bandpass filtering to remove baseline drift and high-frequency electromyography interference, thereby generating a denoised and clean ECG signal time series. The clean ECG signal time series was subjected to multi-scale decomposition based on wavelet transform to extract detail coefficients corresponding to the ventricular depolarization band, and all local energy maxima were identified using an adaptive threshold detection algorithm to generate an initial set of candidate R-wave peak locations. The initial set of candidate R-wave peak locations is subjected to dual screening processing based on amplitude and polarity. Candidates with amplitudes lower than the preset minimum detectable threshold of ECG signals and those with polarity opposite to the dominant heart rhythm are removed to obtain an effective R-wave peak location sequence. Calculate the time interval between two adjacent R-wave peak positions in the effective R-wave peak position sequence to obtain the RR interval sequence; Physiological consistency verification analysis was performed on the RR interval sequence. The R wave peak position corresponding to the interval value whose relative change rate with the previous interval value exceeded the preset allowable range of cardiac chronotropicity was removed from the effective R wave peak position sequence. The signal segment near the removed position was subjected to secondary differential zero-crossing detection to correct the missed and / or false R wave peak positions, and the corrected accurate R wave peak position sequence was obtained. Based on the time coordinates of each R-wave peak position in the corrected precise R-wave peak position sequence, a backward search is performed in the clean ECG signal time series to locate the nearest local minimum of the signal amplitude before each R-wave peak position as the P-wave starting point of the current heartbeat. Simultaneously, the search continues backward to locate the point where the signal amplitude returns to the baseline level and the second derivative approaches zero after each R-wave peak, which is taken as the T-wave termination point of the current heartbeat. The complete signal segment between each precise R-wave peak position and its corresponding P-wave start position and T-wave end position is extracted from the clean ECG signal time series to obtain a single heartbeat waveform unit containing a complete P-QRS-T wave group structure. Arrange all the extracted single heartbeat waveform units in chronological order according to the position of their R-wave peaks to generate a heartbeat waveform unit sequence composed of consecutive single heartbeat waveform units.

[0005] In some possible implementations of the first aspect, the PPG signal time series in the multimodal physiological signal time series is subjected to successive pulse waveform structure segmentation to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units, including: Local maximum point detection processing is performed on the PPG signal time series to identify the locations of all sampling points in the PPG signal time series whose amplitude is greater than the amplitude of the sampling point to its left and the amplitude of the sampling point to its right, and to generate an initial set of candidate pulse wave peak locations. The initial set of candidate pulse wave peak locations is subjected to amplitude filtering processing to remove candidate peak locations whose amplitude is lower than the preset minimum amplitude threshold of detectable PPG signal, thus obtaining the effective pulse wave peak location sequence. Calculate the time interval between two adjacent peaks in the effective pulse wave peak position sequence to obtain the successive pulse interval value sequence. A physiological rationality analysis was performed on the numerical sequence of intervals between successive pulses. The position of the next peak point corresponding to the interval value in the numerical sequence of intervals between successive pulses that differs from the previous interval value from the preset physiological fluctuation range was removed from the effective pulse wave peak point position sequence, resulting in a corrected pulse wave peak point position sequence. Based on the time coordinates of each peak position in the corrected pulse wave peak position sequence, the nearest local minimum of signal amplitude before each peak position is located in the PPG signal time series as the waveform start point of the pulse, and the nearest local minimum of signal amplitude after each peak position is located as the waveform end point of the pulse. The signal segment between each peak position and its corresponding waveform start and end positions is extracted from the PPG signal time series to obtain a complete single pulse waveform unit. All the captured single pulse waveform units are arranged in chronological order of their peak positions to generate a pulse waveform unit sequence composed of consecutive single pulse waveform units.

[0006] In some possible implementations of the first aspect, the evolution trajectory of the heartbeat waveform morphology features is constructed based on the QRS wave characteristics of each heartbeat waveform unit in the heartbeat waveform unit sequence, including: The first sampling point in each heartbeat waveform unit where the QRS wave signal amplitude is lowest and begins to rise is identified as the starting point of that heartbeat waveform unit. The sampling point in each heartbeat waveform unit where the QRS wave signal amplitude is highest is identified as the peak point of that heartbeat waveform unit. The last sampling point in each respiratory waveform unit where the QRS wave signal amplitude is lowest and ends its decline is identified as the ending point of that heartbeat waveform unit. Calculate the time interval between rising points, the time interval between peak points, and the time interval between ending points for adjacent heartbeat waveform units; The rising point time interval, peak point time interval, and termination point time interval are combined and spliced ​​along the feature dimension to form a heartbeat waveform morphological feature evolution trajectory containing rising point time interval information, peak point time interval information, and termination point time interval information.

[0007] In some possible implementations of the first aspect, the evolution trajectory of pulse waveform morphology features is constructed based on the rising slope characteristics, falling steepness characteristics, and relative height characteristics of the diphthoplast notch of each pulse waveform unit in the pulse waveform unit sequence, including: The first sampling point with the lowest signal amplitude in each pulse waveform unit is identified as the waveform start point of that pulse waveform unit; the sampling point with the highest signal amplitude in each pulse waveform unit is identified as the waveform peak point of that pulse waveform unit; and the last sampling point with the lowest signal amplitude in each pulse waveform unit is identified as the waveform end point of that pulse waveform unit. Based on the time coordinates and amplitudes of the waveform's starting point and peak point, calculate the average amplitude change per unit time from the waveform's starting point to its peak point, and use the average amplitude change per unit time as the rising slope characteristic of the pulse waveform unit. Based on the time coordinates and amplitudes of the waveform peak and the waveform termination point, calculate the average amplitude change per unit time from the waveform peak to the waveform termination point, and use the average amplitude change per unit time as the steepness characteristic of the descending branch corresponding to the pulse waveform unit. In the descending phase of each pulse waveform unit, the first local minimum of the signal amplitude is searched after the peak of the waveform, and this local minimum of the signal amplitude is marked as the diabetic wave notch of the pulse waveform unit. Extract the amplitude of the diphthalamic wave notch, calculate the absolute value of the difference between the amplitude of the diphthalamic wave notch and the amplitude of the waveform peak, take the absolute value of the difference as the absolute height value of the diphthalamic wave notch corresponding to the pulse waveform unit, and divide the absolute height value of the diphthalamic wave notch by the amplitude of the waveform peak to obtain the relative height feature of the diphthalamic wave notch corresponding to the pulse waveform unit. Based on the temporal order of each pulse waveform unit in the pulse waveform unit sequence, the rising slope feature corresponding to each pulse waveform unit is arranged sequentially to generate the first feature sub-trajectory of the rising slope feature changing over time; based on the temporal order of each pulse waveform unit in the pulse waveform unit sequence, the descending steepness feature corresponding to each pulse waveform unit is arranged sequentially to generate the second feature sub-trajectory of the descending steepness feature changing over time; based on the temporal order of each pulse waveform unit in the pulse waveform unit sequence, the dicrotic notch relative height feature corresponding to each pulse waveform unit is arranged sequentially to generate the third feature sub-trajectory of the dicrotic notch relative height feature changing over time. The first, second, and third feature sub-trajectories are combined and spliced ​​together along the feature dimension to form a waveform morphology feature evolution trajectory that includes information on the slope evolution of the ascending branch, the steepness evolution of the descending branch, and the relative height evolution of the diphthong notch.

[0008] In some possible implementations of the first aspect, the evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features are synchronously correlated and aligned on the time axis, including: The time coordinates of the QRS wave peaks corresponding to each heartbeat waveform unit are extracted from the evolution trajectory of the heartbeat waveform morphology features to generate a sequence of heartbeat event occurrence times; the time coordinates of the waveform peaks corresponding to each pulse waveform unit are extracted from the evolution trajectory of the pulse waveform morphology features to generate a sequence of pulse event occurrence times. For each pulse event occurrence time point in the pulse event occurrence time point sequence, find the heartbeat event occurrence time point in the heartbeat event occurrence time point sequence that has the shortest time distance from that pulse event occurrence time point, and determine that heartbeat event occurrence time point as the coupled heartbeat event time point corresponding to that pulse event occurrence time point.

[0009] In some possible implementations of the first aspect, the HRV evaluation network model is trained using the HRV data corresponding to the synchronously correlated and aligned evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features, including: A dual-branch spatiotemporal fusion attention network model is constructed as the HRV evaluation network model. The dual-branch spatiotemporal fusion attention network model includes an ECG feature extraction branch, a PPG feature extraction branch, and a spatiotemporal fusion attention module. The ECG feature extraction branch is used to receive and process the evolution trajectory of heartbeat waveform morphological features, the PPG feature extraction branch is used to receive and process the evolution trajectory of pulse waveform morphological features, and the spatiotemporal fusion attention module is used to perform cross-modal correlation fusion of the ECG spatiotemporal features output by the ECG feature extraction branch and the PPG spatiotemporal features output by the PPG feature extraction branch. The first phase of joint pre-training involves inputting the synchronously correlated and aligned evolution trajectory of the heartbeat waveform morphology features into the ECG feature extraction branch and the synchronously correlated and aligned evolution trajectory of the pulse waveform morphology features into the PPG feature extraction branch. The ECG feature extraction branch uses a multi-layer one-dimensional convolutional neural network and a bidirectional gated recurrent unit network to extract spatiotemporal features from the heartbeat waveform morphology feature evolution trajectory, yielding ECG spatiotemporal features. Similarly, the PPG feature extraction branch uses the same multi-layer one-dimensional convolutional neural network and a bidirectional gated recurrent unit network to extract spatiotemporal features from the pulse waveform morphology feature evolution trajectory, yielding PPG spatiotemporal features. Both ECG and PPG spatiotemporal features are then input into the spatiotemporal fusion attention module, utilizing... The multi-head self-attention mechanism and cross-attention mechanism in the spatiotemporal fusion attention module perform feature interaction and weight allocation on ECG spatiotemporal features and PPG spatiotemporal features to generate a fused spatiotemporal feature representation. The fused spatiotemporal feature representation is input into a fully connected regression layer and outputs predicted HRV data. Based on the mean square error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronization association alignment processing, as well as the consistency loss of ECG spatiotemporal features and PPG spatiotemporal features, a joint loss function is constructed. The backpropagation algorithm is used to iteratively update all network parameters of the dual-branch spatiotemporal fusion attention network model until the joint loss function converges to a preset threshold or reaches a preset training round, thus completing the first stage of joint pre-training. The second-stage process involves independent fine-tuning of the PPG feature extraction branch. All network parameters of the ECG feature extraction branch are frozen, or the ECG feature extraction branch is removed from the dual-branch spatiotemporal fusion attention network model. The pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment is input into the PPG feature extraction branch. The multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the PPG feature extraction branch are used to extract spatiotemporal features from the pulse waveform morphology feature evolution trajectory, resulting in PPG spatiotemporal features. These PPG spatiotemporal features are then input into the spatiotemporal fusion attention module, which performs self-attention feature processing on the PPG spatiotemporal features. The process involves enhancing the PPG spatiotemporal feature representation, inputting it into a fully connected regression layer, and outputting predicted HRV data. A fine-tuning loss function is constructed based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing. The network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer are iteratively updated only through backpropagation until the fine-tuning loss function converges to a preset threshold or reaches a preset training epoch. This completes the second stage of independent fine-tuning of the PPG feature extraction branch, resulting in the final trained HRV evaluation network model.

[0010] Among some possible implementations of the first aspect, the method also includes: The IMU signal time series of the target object collected by the wearable device within a preset monitoring period is obtained; among them, the sampling frequencies of the IMU signal time series, ECG signal time series and PPG signal time series are different, but they have a unified time reference timestamp. Based on a unified time reference timestamp, the IMU signal time series, ECG signal time series, and PPG signal time series are resampled or interpolated to put the signal time series with different sampling frequencies onto the same time axis. Then, based on the cross-correlation analysis algorithm, the correlation between the periodic motion peaks in the IMU signal time series and the noise peaks caused by motion in the ECG and PPG signal time series is calculated to automatically compensate for the transmission delay differences of each signal time series. Furthermore, by detecting the time difference between the R-wave cardiac electrical contraction start point in the ECG signal time series and the peak pulse arrival time point in the PPG signal time series, the motion state calculated by the IMU signal time series is combined to dynamically correct the delay fluctuations caused by changes in vascular elasticity in the PPG signal time series, and finally, the synchronization of the IMU signal time series, ECG signal time series, and PPG signal time series is completed. The synchronized IMU signal time series is used as the reference noise signal input, and the synchronized ECG signal time series and PPG signal time series are used as the main input signals. An adaptive filtering algorithm is used to perform motion artifact filtering on the synchronized ECG signal time series and PPG signal time series in order to remove the motion noise components related to the IMU signal time series from the synchronized ECG signal time series and PPG signal time series. The ECG signal time series in the multi-mode physiological signal time series is subjected to successive heartbeat waveform structure segmentation to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units, including: The filtered ECG signal time series is subjected to successive heartbeat waveform structure segmentation to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units. The PPG signal time series in the multimodal physiological signal time series is subjected to successive pulse waveform structure segmentation to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units, including: The filtered PPG signal time series is subjected to successive pulse waveform structure segmentation to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units.

[0011] In some possible implementations of the first aspect, the HRV evaluation network model is trained using the HRV data corresponding to the synchronously correlated and aligned evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features, including: A three-branch spatiotemporal fusion attention network model is constructed as the HRV evaluation network model. The three-branch spatiotemporal fusion attention network model includes an IMU feature extraction branch, an ECG feature extraction branch, a PPG feature extraction branch, and a spatiotemporal fusion attention module. The IMU feature extraction branch is used to receive and process the time series of synchronized IMU signals. The ECG feature extraction branch is used to receive and process the evolution trajectory of heartbeat waveform morphology features after synchronization and alignment. The PPG feature extraction branch is used to receive and process the evolution trajectory of pulse waveform morphology features after synchronization and alignment. The spatiotemporal fusion attention module is used to perform cross-modal correlation fusion on the IMU spatiotemporal features output by the IMU feature extraction branch, the ECG spatiotemporal features output by the ECG feature extraction branch, and the PPG spatiotemporal features output by the PPG feature extraction branch. The first phase of joint pre-training involves inputting the synchronized IMU signal time series into the IMU feature extraction branch, the synchronized and aligned heartbeat waveform morphological feature evolution trajectory into the ECG feature extraction branch, and the synchronized and aligned pulse waveform morphological feature evolution trajectory into the PPG feature extraction branch. Spatiotemporal features are extracted from the input data using multi-layer one-dimensional convolutional neural networks and bidirectional gated recurrent unit networks in each feature extraction branch, yielding IMU spatiotemporal features, ECG spatiotemporal features, and PPG spatiotemporal features, respectively. These features are then input into the spatiotemporal fusion attention module, utilizing spatiotemporal fusion... The multi-head self-attention mechanism and cross-attention mechanism in the attention module perform feature interaction and weight allocation on multimodal spatiotemporal features to generate a fused spatiotemporal feature representation. The fused spatiotemporal feature representation is input into a fully connected regression layer to output predicted HRV data. Based on the mean square error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing, as well as the consistency loss of spatiotemporal features of each branch, a joint loss function is constructed. The backpropagation algorithm is used to iteratively update all network parameters of the three-branch spatiotemporal fusion attention network model until the joint loss function converges to a preset threshold or reaches a preset training round, thus completing the first stage of joint pre-training. The second-stage multi-branch independent fine-tuning process is executed. All network parameters of the ECG and IMU feature extraction branches are frozen, or these branches are removed from the three-branch spatiotemporal fusion attention network model. The pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment is input to the PPG feature extraction branch. The multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the PPG feature extraction branch are used to extract spatiotemporal features from the pulse waveform morphology feature evolution trajectory, obtaining PPG spatiotemporal features. These PPG spatiotemporal features are then input to the spatiotemporal fusion attention module, which is used to process the PPG spatiotemporal features. Self-attention feature enhancement is performed to generate PPG-enhanced spatiotemporal feature representations. These PPG-enhanced spatiotemporal feature representations are then input into a fully connected regression layer, which outputs predicted HRV data. A fine-tuning loss function is constructed based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing. The network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer are iteratively updated only through the backpropagation algorithm until the fine-tuning loss function converges to a preset threshold or reaches a preset number of training rounds. This completes the second stage of independent fine-tuning of the PPG feature extraction branch, resulting in the final trained HRV evaluation network model.

[0012] Secondly, embodiments of this disclosure provide an HRV status assessment device based on PPG data, the device comprising: The acquisition module is used to acquire the time series of multimodal physiological signals of the target object collected by the wearable device within a preset monitoring period; The segmentation module is used to perform successive heartbeat waveform structure segmentation on the ECG signal time series in the multi-mode physiological signal time series to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units; and to perform successive pulse waveform structure segmentation on the PPG signal time series in the multi-mode physiological signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units. The module is used to construct the evolution trajectory of heartbeat waveform morphology features based on the QRS wave features of each heartbeat waveform unit in the heartbeat waveform unit sequence; and to construct the evolution trajectory of pulse waveform morphology features based on the slope features of the ascending branch, the steepness features of the descending branch, and the relative height features of the diphthong notch of each pulse waveform unit in the pulse waveform unit sequence. The alignment module is used to synchronously align the evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features on the time axis; and to calculate the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous alignment. The training module is used to train the HRV evaluation network model using the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory and pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment processing. The evaluation module is used to acquire the PPG signal time series of the target object collected by the wearable device in real time, construct the pulse waveform morphology feature evolution trajectory in real time based on the acquired PPG signal time series, input the constructed pulse waveform morphology feature evolution trajectory into the trained HRV evaluation network model, and output the corresponding HRV data.

[0013] Thirdly, embodiments of this disclosure provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0014] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described above.

[0015] In summary, the embodiments of this disclosure, by constructing and synchronously associating the evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features, fully utilize the precision of ECG signals in the timing of cardiac electrical activity and the richness of PPG signals in hemodynamic morphology, achieving complementary advantages and deep alignment of multimodal data. By using this high-precision synchronous trajectory data to train the HRV assessment network model, the model can learn the complex nonlinear mapping relationship from microscopic morphological changes of PPG waveforms to HRV indicators. Thus, with only low-power real-time acquired PPG signals as input, it can output HRV data with measurement accuracy close to that of ECG signals.

[0016] This not only completely eliminates the continuous reliance on high-power ECG circuits and significantly extends the battery life of wearable devices, meeting users' needs for all-weather, long-term health monitoring, but more importantly, by utilizing the alignment and complementarity effects of multimodal data, it greatly suppresses the interference of ambient light, motion artifacts, and individual differences on PPG signals, improving the reliability of assessment in scenarios with poor waveform morphology stability. This results in high accuracy of HRV analysis based on PPG signals, which can accurately reflect the true state of autonomic nervous system regulation, providing a technical solution for clinical disease early warning and daily health management that combines low cost, low power consumption, and high confidence.

[0017] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0018] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart of an HRV status assessment method based on PPG data provided in this disclosure is shown. Figure 2 A structural diagram of an HRV status assessment device based on PPG data provided in an embodiment of this disclosure is shown. Figure 3 A structural diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0020] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0021] To address the problems in the background art, this disclosure provides a method and apparatus for HRV status assessment based on PPG data. By constructing and synchronously associating the evolution trajectory of heartbeat waveform morphology features with the evolution trajectory of pulse waveform morphology features, it fully utilizes the accuracy of ECG signals in cardiac electrical activity timing and the richness of PPG signals in hemodynamic morphology, achieving complementary advantages and deep alignment of multimodal data. Using this high-precision synchronous trajectory data to train an HRV assessment network model, the model can learn the complex nonlinear mapping relationship from microscopic morphological changes of PPG waveforms to HRV indicators. Thus, with only low-power real-time acquired PPG signals as input, it can output HRV data with measurement accuracy close to that of ECG signals.

[0022] The following detailed description, with reference to the accompanying drawings and specific embodiments, illustrates a method and apparatus for HRV status assessment based on PPG data provided in this disclosure.

[0023] Figure 1 A flowchart of a HRV status assessment method based on PPG data provided in this disclosure is shown, as follows: Figure 1 As shown, method 100 may include the following steps: S110, acquire the time series of multi-mode physiological signals of the target object collected by wearable devices within a preset monitoring period.

[0024] The wearable device has built-in ECG and PPG sensors, and the multi-mode physiological signal time series it collects include ECG signal time series and PPG signal time series.

[0025] As an example, within a preset monitoring period, such as a continuous 24-hour time window from the start time T_start to the end time T_end, the wearable device can continuously collect PPG and ECG signals of the target object at a fixed high sampling frequency.

[0026] Specifically, the ECG sensor detects the weak potential difference (typical amplitude 0.5–5 mV) generated by cardiac electrical activity through surface electrodes at a sampling frequency Fs_ecg. The analog voltage signal on the skin surface is converted into a digital ECG signal time series E_ecg by an analog-to-digital converter. Each sampling point e_ecg(t) in this series contains a high-precision timestamp t (with millisecond precision) and an amplitude a_ecg (in microvolts, which corresponds to the actual potential difference after calibration) reflecting cardiac electrophysiological activity. Simultaneously, the PPG sensor emits a light beam of a specific wavelength (selectable from green light 525nm, red light 660nm, and infrared light 940nm) at a sampling frequency Fs_ppg to irradiate the skin tissue and receives changes in the intensity of reflected light (modulated by vascular volume pulsation). The analog light intensity signal is converted into a digital PPG signal time series P_ppg by an analog-to-digital converter. Each sampling point p_ppg(t) in this series contains a high-precision timestamp t (with millisecond precision) and an amplitude a_ppg reflecting the change in vascular volume (in units of relative light intensity, normalized to eliminate the influence of individual skin differences).

[0027] It is worth noting that after the wearable device collects the multi-mode physiological signal time series (ECG signal time series and PPG signal time series), the data can be retrieved from the wearable device. After retrieval, timestamp alignment verification is performed on each signal time series to check for data packet loss or timestamp misalignment. For individual missing sampling points, an interpolation method based on local weighted regression is used to fill in the missing points. That is, the weighted average value is calculated based on the amplitudes of several valid sampling points before and after the missing point as the estimated amplitude of the missing point, thereby ensuring that each signal time series is strictly equally spaced and complete on the time axis.

[0028] S120, the ECG signal time series in the multi-mode physiological signal time series is processed by successive heartbeat waveform structure segmentation to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units; and the PPG signal time series in the multi-mode physiological signal time series is processed by successive pulse waveform structure segmentation to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units.

[0029] In some embodiments, the above-described process of segmenting the ECG signal time series in a multimodal physiological signal time series into successive heartbeat waveform structures to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units includes: (1) Bandpass filtering is performed on the ECG signal time series to filter out baseline drift and high-frequency electromyography interference, and a clean ECG signal time series after noise reduction is generated.

[0030] Specifically, a digital bandpass filter is designed with a passband frequency range covering the main energy distribution range of the ECG signal (e.g., 0.5Hz to 40Hz) to suppress low-frequency baseline drift and high-frequency electromyographic noise. The filter is applied to the ECG signal time series E_ecg, and the amplitude a_ecg of e_ecg(t) at each sampling point is convolved or recursively calculated to obtain the filtered new amplitude a_clean. All filtered sampling points are reassembled in their original time order to form the denoised clean ECG signal time series E_ecg_clean. This series retains key physiological features such as the P wave, QRS complex (also known as the QRS group), and T wave, while significantly reducing interference from non-physiological noise.

[0031] (2) Perform multi-scale decomposition processing based on wavelet transform on the clean ECG signal time series, extract the detail coefficients corresponding to the ventricular depolarization frequency band, and use the adaptive threshold detection algorithm to identify all local energy maxima points to generate an initial R-wave peak candidate point location set.

[0032] Specifically, a discrete wavelet transform is performed on the clean ECG signal time series E_ecg_clean. A wavelet basis function (such as db4 or sym4) suitable for the ECG signal characteristics is selected to decompose the signal into detail coefficients at multiple scales. A specific scale detail coefficient sequence D_qrs corresponding to the main frequency range of ventricular depolarization (QRS wave) is selected. The local energy of this detail coefficient sequence within a sliding window is calculated, and an adaptive threshold Th_wavelet is set that dynamically adjusts with the signal noise level. The detail coefficient sequence is traversed to identify all points whose amplitude exceeds Th_wavelet and is a local maximum in its neighborhood; these points correspond to potential R-wave peak locations. The timestamps of these points are recorded in a dynamic list, forming the initial R-wave peak candidate location set C_r_init. Due to the robustness of wavelet transform to noise, this set initially filters out most of the true R-wave locations, but may still contain some false candidate points caused by high-energy artifacts.

[0033] (3) Perform dual screening of amplitude and polarity on the initial set of candidate R wave peak locations, and remove candidate points whose amplitude is lower than the preset minimum detectable threshold of ECG signal and whose polarity is opposite to the dominant heart rhythm to obtain the effective R wave peak location sequence.

[0034] Specifically, a minimum detectable amplitude threshold Th_ecg_min for the ECG signal is obtained from the configuration parameters to distinguish between the actual ventricular depolarization signal and residual noise. Simultaneously, the polarity direction of most candidate points in the initial R-wave peak candidate point location set C_r_init is analyzed to determine the polarity standard of the dominant rhythm. For each candidate point location c_j in the set C_r_init, its corresponding original amplitude a is obtained from the clean ECG signal time series E_ecg_clean. ecg First, it is determined whether the absolute value of a_ecg_j is less than Th_ecg_min. If so, it is considered weak noise interference and is removed. Second, it is determined whether the polarity of a_ecg_j is opposite to the polarity of the dominant heart rhythm. If so, it is determined as a polarity artifact and is removed. After these two screenings, the remaining candidate points are rearranged in chronological order to form the effective R-wave peak position sequence R_valid. Each point in this sequence is a reliable R-wave position that conforms to physiological laws in terms of amplitude intensity and polarity characteristics.

[0035] (4) Calculate the time interval between two adjacent R-wave peak positions in the effective R-wave peak position sequence to obtain the RR interval sequence.

[0036] Specifically, the valid R-wave peak position sequence R_valid is processed. The R-wave peaks in this sequence are arranged in ascending order of time as r1, r2, r3, ..., rM. For each pair of adjacent R-wave peaks r_k and r_{k+1}, their corresponding high-precision timestamps t_k and t_{k+1} in the clean ECG signal time sequence E_ecg_clean are extracted. The time difference between the two is calculated as Δt_k = t_{k+1} - t_k, which is the RR interval between the k-th and (k+1)-th heartbeats. All calculated RR intervals are arranged according to their corresponding heartbeat order, forming a one-dimensional numerical sequence of length M-1, i.e., the RR interval sequence I_rr. Each element I_rr_k in this sequence represents the time length between two adjacent heartbeats, in milliseconds, reflecting the instantaneous changes in heart rhythm.

[0037] (5) Perform physiological consistency verification analysis on the RR interval sequence, remove the position of the next R wave peak corresponding to the interval value in the RR interval sequence whose relative change rate with the previous interval value exceeds the preset allowable range of cardiac chronotropicity from the effective R wave peak position sequence, and perform secondary differential zero-crossing detection on the signal segment near the removed position to correct the missed and / or false R wave peak positions, so as to obtain the corrected accurate R wave peak position sequence.

[0038] Specifically, under physiological conditions, the RR interval changes in normal sinus rhythm are gradual, without abrupt, drastic jumps. A permissible range for cardiac chronotropic changes is set, typically defined as the maximum tolerable value of the relative change rate between adjacent intervals, Ratio_max. Iterating through the RR interval sequence I_rr, starting from the second element, for the k-th interval value I_rr_k, calculate its relative change rate with the previous interval value I_rr_{k-1}, Rate_k = |I_rr_k - I_rr_{k-1}| / I_rr_{k-1}. If Rate_k exceeds the preset threshold Ratio_max, the k-th interval is considered abnormal, and its corresponding subsequent R-wave peak point r_{k+1} may be a false positive or a missed positive. This abnormal point r_{k+1} is marked for removal. Subsequently, for the clean ECG signal time series E_ecg_clean segment within the time window near the rejected point (e.g., 200 ms before and after), a second derivative is performed to find the zero-crossing point of the second derivative to accurately locate the true R-wave peaks or false detections that may have been missed. If a new peak matching the characteristics is detected through the second derivative, it is added to the sequence; if it is confirmed as a false detection, it is completely removed. After all verification and correction are completed, the remaining peaks are arranged in chronological order to form the corrected accurate R-wave peak location sequence R_corrected, which highly accurately reflects continuous ventricular contraction events.

[0039] (6) Based on the time coordinates of each R-wave peak position in the corrected precise R-wave peak position sequence, backtrack in the clean ECG signal time series to locate the nearest local minimum of the signal amplitude before each R-wave peak position as the P-wave starting point of the current heartbeat.

[0040] Specifically, for each R-wave peak r_n in the corrected R-wave peak position sequence R_corrected, the starting boundary of its corresponding heartbeat waveform needs to be determined. Starting from the time coordinate of r_n, a backtracking search is performed to the left (in the earlier time direction) within the clean ECG signal time sequence E_ecg_clean. During the search, amplitudes are compared point by point to find the first point that satisfies the local minimum condition, i.e., the amplitude of this point is simultaneously less than the amplitudes of its left and right adjacent points, and this point is located within the expected P-wave time window (usually 120-200 milliseconds before the R-wave). This local minimum point is determined as the P-wave starting point S_p_n of the current heartbeat, marking the start of atrial depolarization. If the T-wave termination point of the previous heartbeat has been reached before a clear local minimum is found, the lowest potential point between the two waves is taken as the starting point.

[0041] (7) Simultaneously, track and search backwards to locate the point where the signal amplitude returns to the baseline level and the second derivative approaches zero after each R-wave peak position, and take it as the T-wave termination point of the current heartbeat.

[0042] Specifically, starting from the time coordinate of the same R-wave peak point r_n, a tracing search is performed to the right (in a later time direction) within the clean ECG signal time series E_ecg_clean to determine the termination boundary of the heartbeat waveform. The search goal is to find the position where the T wave ends and returns to the baseline. The specific criteria are: the signal amplitude a_ecg stabilizes near the baseline level (i.e., the rate of change of amplitude is close to zero), and the second derivative of the signal approaches zero, indicating that the waveform curvature disappears and enters the isoelectric line stage. The first point that meets the above conditions is determined as the T-wave termination point E_t_n of the current heartbeat, marking the end of this ventricular repolarization.

[0043] (8) Extract the complete signal segment between each precise R wave peak position and its corresponding P wave start position and T wave end position from the clean ECG signal time series to obtain a single heartbeat waveform unit containing a complete P-QRS-T wave group structure.

[0044] Specifically, for the nth heartbeat, using the time indices of the determined P-wave start point S_p_n and T-wave end point E_t_n, all consecutive sampling points (including endpoints) from S_p_n to E_t_n are extracted from the clean sequence E_ecg_clean. This continuous signal segment completely records the entire process from the start of atrial depolarization (P wave), through atrioventricular conduction and ventricular depolarization (QRS wave), until the end of ventricular repolarization (T wave), containing complete morphological information of cardiac electrophysiological activity, and constituting an independent single heartbeat waveform unit B_beat_ecg_n.

[0045] (9) Arrange all the extracted single heartbeat waveform units in chronological order of their R-wave peak positions to generate a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units.

[0046] Specifically, all extracted single-beat waveform units, namely B_beat_ecg_1, B_beat_ecg_2, ..., B_beat_ecg_N, are arranged in ascending order of their corresponding precise R-wave peak timestamps r_n, forming an ordered list. This list is the heartbeat waveform unit sequence B_beat_seq_ecg. Each element B_beat_ecg_n in the sequence is a one-dimensional signal segment of varying length (because the duration of each cardiac cycle is different), representing a complete cardiac electrical activity. The entire sequence comprehensively covers all cardiac pulsations within the monitoring period.

[0047] In some embodiments, the above-described process of successively segmenting the PPG signal time series in the multimodal physiological signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units includes: (1) Perform local maximum point detection processing on the PPG signal time series to identify the sampling point positions in the PPG signal time series whose amplitude is greater than the amplitude of the adjacent sampling point to the left and the amplitude of the adjacent sampling point to the right, and generate an initial set of candidate pulse wave peak positions.

[0048] Specifically, each internal sampling point in the PPG signal time series P_ppg is traversed, excluding the first and last sampling points of the sequence. For any sampling point p_ppg(t_k), the amplitude a_k-1} of its left adjacent sampling point p_ppg(t_{k-1}) and the amplitude a_k+1} of its right adjacent sampling point p_ppg(t_{k+1}) are extracted. By comparison logic, if both a_k and a_k are satisfied, the timestamp t_k of the current sampling point is determined as a local maximum point, which is a candidate position of the pulse wave peak. The time indices of all candidate points that meet this condition are recorded in a dynamic array, which constitutes the initial set of candidate pulse wave peak positions C_peak_init. Since the original signal may contain high-frequency noise or baseline drift, this set may contain some spurious peaks caused by non-physiological factors.

[0049] (2) Perform amplitude screening on the initial pulse wave peak candidate point position set, remove the peak candidate point positions in the initial pulse wave peak candidate point position set whose amplitude is lower than the preset minimum amplitude threshold of the detectable PPG signal, and obtain the effective pulse wave peak point position sequence.

[0050] Specifically, a minimum amplitude threshold Th_ppg_min for detectable PPG signals is read from configuration parameters or adaptively calculated based on signal quality. This threshold is used to distinguish between genuine pulse signals and weak noise interference. Each candidate point position c_i in the initial candidate point set C_peak_init is traversed, and the corresponding amplitude a_i in the original signal P_ppg is obtained. If a_i is less than Th_ppg_min, the candidate point is determined to be possibly caused by electromagnetic interference or sensor jitter and is removed from the set. All candidate points retained after amplitude filtering are reordered according to their chronological order to form an ordered list, namely the effective pulse wave peak position sequence P_peak_valid. Each point in this sequence is a reliable peak position that meets the physiological signal intensity requirements in terms of amplitude.

[0051] (3) Calculate the time interval between two adjacent peaks in the effective pulse wave peak position sequence to obtain the successive pulse interval value sequence.

[0052] Specifically, the effective pulse wave peak position sequence P_peak_valid is processed. The peaks in this sequence are arranged in ascending order of time as p1, p2, p3, ..., pN. For each pair of adjacent peaks p_k and p_{k+1}, their time coordinate difference is calculated, i.e., the timestamp of p_k is subtracted from the timestamp of p_{k+1}. The resulting time difference is the pulse interval between the k-th pulse and the (k+1)-th pulse. All the calculated pulse intervals are arranged according to their corresponding pulse order to form a one-dimensional numerical sequence of length N-1, namely the successive pulse interval numerical sequence I_rri. Each element I_rri_k in this sequence represents the time length between two adjacent pulses, in milliseconds.

[0053] (4) Perform physiological rationality analysis on the interval value sequence between successive pulses, and remove the position of the next peak point corresponding to the interval value in the interval value sequence between successive pulses that exceeds the preset physiological fluctuation range from the effective pulse wave peak point position sequence to obtain the corrected pulse wave peak point position sequence.

[0054] Specifically, under physiological conditions, the variation in adjacent pulse intervals is usually within a certain range and will not result in drastic abrupt changes. A permissible range of physiological fluctuations is set, which can be dynamically generated based on the statistical characteristics of the current sequence. For example, using the median of the current pulse interval sequence as a benchmark, the maximum permissible percentage deviation from the adjacent intervals is calculated. In practice, the sequence of successive pulse interval values ​​I_rri is traversed. Starting from the second interval element, for the k-th interval value I_rri_k, it is compared with the previous interval value I_rri_{k-1}, and the absolute difference ΔI_k = |I_rri_k - I_rri_{k-1}| is calculated. If ΔI_k exceeds the preset maximum permissible fluctuation threshold Th_rri_delta, then the subsequent peak point p_{k+1} corresponding to the k-th interval value I_rri_k is considered to be due to false detection (such as misdetecting a dicrotic wave as the main wave) or abnormal pulsation in arrhythmia, which does not conform to the normal sinus rhythm. The abnormal peak point p_{k+1} is marked as to be removed in the valid pulse wave peak point position sequence P_peak_valid. After traversing the entire interval sequence, all marked peak points to be removed are removed from the sequence, and the remaining peak points are arranged in their original order to form the corrected pulse wave peak point position sequence P_peak_corrected. The peak points in this sequence are considered to more accurately reflect continuous and physiologically reasonable ventricular contraction events.

[0055] (5) Based on the time coordinates of each peak position in the corrected pulse wave peak position sequence, locate the nearest local minimum signal amplitude point before each peak position in the PPG signal time sequence as the waveform start point of the pulse, and locate the nearest local minimum signal amplitude point after each peak position as the waveform end point of the pulse.

[0056] Specifically, for each peak point p_m in the corrected pulse wave peak point position sequence P_peak_corrected, it is necessary to determine the complete boundary of the pulse waveform to which it belongs. Starting from the position of p_m, search to the left (earlier in time) in the PPG signal time series P_ppg to find the first local minimum point. The search process compares the amplitudes point by point. When it is detected that the amplitude of a certain point s is simultaneously less than the amplitudes of its adjacent points on the left and right, that is, a_s < a_{s - 1} and a_s < a_{s + 1}, then this point s is determined as the waveform starting point S_beat_m of this pulse, which corresponds to the end of ventricular diastole. Then, starting from the position of p_m, search to the right (later in time) to find the first local minimum point, that is, the point e that satisfies the condition a_e < a_{e - 1} and a_e < a_{e + 1}, and this point e is determined as the waveform ending point E_beat_m of this pulse, which corresponds to the start of the diastole of the next cardiac cycle. In this way, each pulse wave peak obtains accurate starting and ending boundaries.

[0057] (6) Extract the signal segment between each peak point position and its corresponding waveform starting point position and waveform ending point position from the PPG signal time series to obtain a complete single - pulse waveform unit.

[0058] Specifically, for the m - th pulse, all consecutive sampling points (including these two endpoints) between the waveform starting point S_beat_m and the waveform ending point E_beat_m are completely extracted from the original signal P_ppg. This continuous signal segment contains the complete morphological information from the end of diastole, through the ascending branch formed during the rapid ejection period, the systolic peak, the descending branch formed during the reduced ejection period, and the dicrotic notch formed by the closure of the aortic valve, until the start of the next diastole, and constitutes an independent single - pulse waveform unit B_beat_m.

[0059] (7) Arrange all the extracted single - pulse waveform units in the chronological order of their peak point positions to generate a pulse waveform unit sequence composed of consecutive single - pulse waveform units.

[0060] Specifically, all the extracted single - pulse waveform units, that is, B_beat_1, B_beat_2,..., B_beat_M, are arranged in ascending order of the time stamps of their corresponding peak points p_m to form an ordered list, which is the pulse waveform unit sequence B_beat_seq. Each element B_beat_m in this sequence is a one - dimensional signal segment with a possibly different length (because the duration of each pulse cycle is different), representing a complete pulse waveform, and the entire sequence completely covers all the pulse activities within the monitoring period.

[0061] S130, construct the evolution trajectory of heartbeat waveform morphology features based on the QRS wave characteristics of each heartbeat waveform unit in the heartbeat waveform unit sequence; and construct the evolution trajectory of pulse waveform morphology features based on the slope characteristics of the ascending branch, the steepness characteristics of the descending branch, and the relative height characteristics of the diphthong notch of each pulse waveform unit in the pulse waveform unit sequence.

[0062] In some embodiments, the above-mentioned construction of the heartbeat waveform morphological feature evolution trajectory based on the QRS wave features of each heartbeat waveform unit in the heartbeat waveform unit sequence includes: (1) Locate the first sampling point in each heartbeat waveform unit where the QRS wave signal amplitude is the lowest and begins to rise as the starting point of the heartbeat waveform unit, locate the sampling point in each heartbeat waveform unit where the QRS wave signal amplitude is the highest as the peak point of the heartbeat waveform unit, and locate the last sampling point in each respiratory waveform unit where the QRS wave signal amplitude is the lowest and ends its decline as the termination point of the heartbeat waveform unit.

[0063] Specifically, for each independent single heartbeat waveform unit B_beat_ecg_n in the heartbeat waveform unit sequence B_beat_seq_ecg, within its QRS wave range, the first sampling point where the amplitude aecg reaches its lowest state and then immediately enters the rising phase is located, and this point is defined as the starting point of the unit. Then, the sampling point where the amplitude aecg reaches the global maximum value is located is taken as the peak point. Finally, the last sampling moment when the amplitude aecg is at its lowest level after completing the falling process is determined as the termination point.

[0064] (2) Calculate the time interval between rising points, peak points, and termination points of adjacent heartbeat waveform units.

[0065] Specifically, the time difference between the starting points of adjacent heartbeat waveform units is calculated to obtain the rising point time interval, the time difference between the peak points of adjacent heartbeat waveform units is calculated to obtain the peak point time interval, and the time difference between the ending points of adjacent heartbeat waveform units is calculated to obtain the ending point time interval, thereby quantifying the evolution rate of the heartbeat waveform morphology in time.

[0066] (3) Combine and splice the rising point time interval, peak point time interval, and termination point time interval in the feature dimension to form a heartbeat waveform morphological feature evolution trajectory containing rising point time interval information, peak point time interval information, and termination point time interval information.

[0067] Specifically, the rise point time interval, peak point time interval, and termination point time interval are regarded as three independent feature dimensions. According to the time evolution order, these interval data of all adjacent units in the heartbeat waveform unit sequence B_beat_seq_ecg are vector-combined and spliced ​​to construct a multidimensional heartbeat waveform morphological feature evolution trajectory that can simultaneously reflect the dynamic change law of the start, peak and termination stages. This trajectory fully describes the morphological stability and variation trend of cardiac electrical activity in the continuous heartbeat cycle.

[0068] In some embodiments, the above-mentioned construction of the pulse waveform morphological feature evolution trajectory based on the rising slope feature, falling steepness feature, and relative height feature of the diphthoplast notch of each pulse waveform unit in the pulse waveform unit sequence includes: (1) Locate the first sampling point with the lowest signal amplitude in each pulse waveform unit as the waveform start point of the pulse waveform unit, locate the sampling point with the highest signal amplitude in each pulse waveform unit as the waveform peak point of the pulse waveform unit, and locate the last sampling point with the lowest signal amplitude in each pulse waveform unit as the waveform end point of the pulse waveform unit.

[0069] Specifically, for each unit B_beat_m in the pulse waveform unit sequence B_beat_seq, the waveform start point S_beat_m (i.e., the first local minimum point within the unit) and waveform end point E_beat_m (i.e., the last local minimum point within the unit) of that unit are directly used. Simultaneously, the waveform peak point P_beat_m of that unit is the main peak point of the pulse, located at the position of maximum signal amplitude within the unit.

[0070] (2) Based on the time coordinates and amplitude of the waveform starting point and the time coordinates and amplitude of the waveform peak point, calculate the average amplitude change per unit time from the waveform starting point to the waveform peak point, and use the average amplitude change per unit time as the rising slope characteristic of the pulse waveform unit.

[0071] Specifically, the timestamp t_s and amplitude a_s of the starting point S_beat_m of the m-th pulse waveform unit, and the timestamp t_p and amplitude a_p of the peak point P_beat_m are extracted. To characterize the rate of pressure rise from the end of diastole to the peak of systole, the time increment Δt_up (t_p minus t_s) and the amplitude increment Δa_up (a_p minus a_s) are calculated. Then, the ratio of the amplitude increment to the time increment is defined as the rising slope characteristic K_up_m, i.e., K_up_m equals Δa_up divided by Δt_up. This ratio represents the average rate of pressure rise during the initial stage of ventricular ejection, and its numerical change reflects the change in the strength of myocardial contractility.

[0072] (3) Based on the time coordinates and amplitudes of the waveform peak and the waveform termination point, calculate the average amplitude change per unit time from the waveform peak to the waveform termination point, and use the average amplitude change per unit time as the steep descending branch characteristic of the pulse waveform unit.

[0073] Specifically, the timestamp t_p and amplitude a_p of the peak point P_beat_m, and the timestamp t_e and amplitude a_e of the termination point E_beat_m are extracted. To characterize the rate of pressure drop after contraction, the time increment Δt_down (t_e minus t_p) and the amplitude decrease Δa_down (a_p minus a_e) are calculated. Then, the ratio of amplitude decrease to time increment is defined as the descending limb steepness characteristic K_down_m, i.e., K_down_m equals Δa_down divided by Δt_down. This ratio reflects peripheral vascular resistance, aortic compliance, and the attenuation characteristics of the pulse wave in peripheral arteries, and its changes are closely related to the vascular elasticity.

[0074] (4) In the descending phase of each pulse waveform unit, search for the first local minimum point of signal amplitude after the peak point of the waveform, and mark the local minimum point of signal amplitude as the diphthorax notch point of the pulse waveform unit.

[0075] Specifically, the search range is limited to the descending limb of the m-th pulse waveform unit, i.e., from the peak point P_beat_m to the termination point E_beat_m. Within this local signal segment, a local minimum amplitude point is detected point by point, i.e., the amplitude of this point is simultaneously less than the amplitudes of its left and right adjacent points. The first local minimum point that meets the conditions is the dicrotic wave notch D_beat_m. This point corresponds to the notch formed on the pulse wave due to the brief rise in intraaortic pressure caused by the closure of the aortic valve after systole, and is a key marker for assessing vascular elasticity.

[0076] (5) Extract the amplitude of the diphthrombectomy notch, calculate the absolute value of the difference between the amplitude of the diphthrombectomy notch and the amplitude of the waveform peak, take the absolute value of the difference as the absolute height value of the diphthrombectomy notch corresponding to the pulse waveform unit, and divide the absolute height value of the diphthrombectomy notch by the amplitude of the waveform peak to obtain the relative height feature of the diphthrombectomy notch corresponding to the pulse waveform unit.

[0077] Specifically, the amplitude a_d of the dicrotic wave notch point D_beat_m is extracted. To quantify the depth of the dicrotic wave, the absolute difference between the peak amplitude and the notch amplitude is first calculated, i.e., the absolute height value H_abs_m of the dicrotic wave notch, which is equal to the absolute value of a_p minus a_d. Since there are significant differences in pulse wave amplitude between individuals, the absolute height value is difficult to use directly for comparison. Therefore, the absolute height value and the peak amplitude are further normalized, and the relative height characteristic H_rel_m of the dicrotic wave notch is calculated, i.e., H_abs_m divided by a_p. This ratio eliminates the influence of signal gain differences between individuals and more essentially reflects the elastic recoil capacity of the large artery.

[0078] (6) Arrange the rising slope features of each pulse waveform unit in the pulse waveform unit sequence according to the time sequence to generate the first feature sub-trajectory of the rising slope features changing with time; arrange the falling steep features of each pulse waveform unit in the pulse waveform unit sequence according to the time sequence to generate the second feature sub-trajectory of the falling steep features changing with time; arrange the diphthong notch relative height features of each pulse waveform unit in the pulse waveform unit sequence according to the time sequence to generate the third feature sub-trajectory of the diphthong notch relative height features changing with time.

[0079] Specifically, for all M units in the pulse waveform unit sequence, the upward slope feature K_up_m calculated for each unit is organized into a one-dimensional vector T_up according to the unit index m from 1 to M. The m-th element of the vector is the upward slope of the m-th pulse. This vector T_up constitutes the first feature sub-trajectory, which describes the evolution of ventricular systolic force throughout the monitoring period. Similarly, the downward steepness feature K_down_m of each unit is organized into a one-dimensional vector T_down according to the index order, forming the second feature sub-trajectory, which describes the dynamic changes in vascular resistance characteristics. The dicrotic notch relative height feature H_rel_m of each unit is organized into a one-dimensional vector T_dicrotic according to the index order, forming the third feature sub-trajectory, which describes the beat-by-beat evolution of vascular elasticity.

[0080] (7) Combine and splice the first feature sub-trajectory, the second feature sub-trajectory and the third feature sub-trajectory in the feature dimension to form a waveform morphology feature evolution trajectory that includes the evolution information of the slope of the rising branch, the evolution information of the steepness of the falling branch and the evolution information of the relative height of the diphtheria wave notch.

[0081] Specifically, three one-dimensional vectors of length M, T_up, T_down, and T_dicrotic, are combined. This combination is achieved by constructing a two-dimensional matrix H_beat_traj, which has M rows and 3 columns. The m-th row corresponds to the m-th pulse. The first column of this row stores the m-th value of T_up, the second column stores the m-th value of T_down, and the third column stores the m-th value of T_dicrotic. This M-row, 3-column matrix represents the evolution trajectory of the pulse waveform morphology. It not only records the morphological characteristics of each pulse but, more importantly, preserves the temporal relationship of the evolution of these characteristics with the pulse sequence, arranged in the row order of the matrix.

[0082] S140, synchronize and align the evolution trajectory of heartbeat waveform morphology features and pulse waveform morphology features on the time axis; and calculate the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronization and alignment.

[0083] In some embodiments, the above-mentioned synchronous correlation and alignment processing of the evolution trajectory of heartbeat waveform morphological features and pulse waveform morphological features on the time axis includes: (1) Extract the time coordinates of the QRS wave peaks corresponding to each heartbeat waveform unit from the evolution trajectory of the heartbeat waveform morphology features to generate a sequence of heartbeat event occurrence time points; extract the time coordinates of the waveform peaks corresponding to each pulse waveform unit from the evolution trajectory of the pulse waveform morphology features to generate a sequence of pulse event occurrence time points.

[0084] Specifically, the time coordinates of the QRS wave peaks corresponding to each heartbeat waveform unit are extracted from the evolution trajectory of the heartbeat waveform morphology features, denoted as t_R_i, and a sequence of heartbeat event occurrence time points is generated based on this, denoted as T_heart={t_R_1, t_R_2, ..., t_R_n}; the time coordinates of the waveform peaks corresponding to each pulse waveform unit are extracted from the evolution trajectory of the pulse waveform morphology features, denoted as t_P_j, and a sequence of pulse event occurrence time points is generated based on this, denoted as T_pulse={t_P_1, t_P_2, ..., t_P_m}.

[0085] (2) For each pulse event occurrence time point in the pulse event occurrence time point sequence, find the heartbeat event occurrence time point with the shortest time distance from the pulse event occurrence time point in the heartbeat event occurrence time point sequence, and determine the heartbeat event occurrence time point as the coupled heartbeat event time point corresponding to the pulse event occurrence time point.

[0086] Specifically, for any given pulse event occurrence time point \(t_{P_j}\), traverse all time points \(t_{R_i}\) in the heart beat event occurrence time point sequence \(T_{heart}\), calculate the absolute value of the time difference between the two, denoted as \(\Delta t_{ij} = |t_{P_j}-t_{R_i}|\). Considering the physiological causal constraint, that is, the pulse wave must lag behind the corresponding electrocardiogram wave, the search range can be further limited to the candidate set where \(t_{R_i}<t_{P_j}\) and the time difference is within a reasonable physiological interval (usually 50 milliseconds to 400 milliseconds, specifically depending on the sensor position). Within this constraint set, select the heart beat event occurrence time point that makes \(\Delta t_{ij}\) the smallest, denoted as \(t_{R_k}\), that is, satisfying the condition \(t_{R_k}=\arg\min|t_{P_j}-t_{R_i}|\), and officially determine it as the coupled heart beat event time point corresponding to the pulse event occurrence time point \(t_{P_j}\). This process essentially establishes a mapping function from mechanical pulsation back to electrical trigger in the time domain, denoted as \(f: T_{pulse}\to T_{heart}\), which not only eliminates the matching ambiguity caused by the beat-to-beat interval fluctuation, but also effectively eliminates the false peaks caused by possible artifacts. Through this point-by-point matching mechanism, a set of one-to-one corresponding coupled event pairs is generated, expressed as \(\{(t_{R_k1}, t_{P_1}), (t_{R_k2}, t_{P_2}),...\}\), and these coupled point pairs not only synchronize the time axes of the two heterogeneous signals, but also directly contain the pulse conduction time information of each heart beat cycle.

[0087] In some embodiments, calculating the HRV data corresponding to the evolution trajectory of the morphological characteristics of the heart beat waveform after the above-mentioned synchronous correlation alignment processing includes: (1) Based on the evolution trajectory of the morphological characteristics of the heart beat waveform after synchronous correlation alignment processing, use the adaptive threshold detection algorithm combined with the wavelet transform modulus maximum method to accurately locate the R wave peak points of the QRS wave corresponding to each heart beat waveform unit from the trajectory, extract the accurate time coordinates of each R wave peak point, obtain the R wave peak point time sequence, and calculate the time interval between adjacent R wave peak points according to the R wave peak point time sequence to obtain the RR interval sequence.

[0088] (2) Perform time-domain statistical analysis on the RR interval sequence, calculate the standard deviation SDNN of all normal sinus rhythm RR intervals, the root mean square RMSSD of the difference between adjacent RR intervals, and the percentage pNN50 of the difference between adjacent RR intervals greater than 50 milliseconds, in order to quantify the instantaneous regulatory ability of the autonomic nervous system; at the same time, use the fast Fourier transform (FFT) or autoregressive model (AR) method to convert the RR interval sequence to the frequency domain, and calculate the very low frequency power (VLF), low frequency power (LF), high frequency power (HF) and their normalized ratios LF_norm, HF_norm and LF / HF ratio, respectively, in order to assess the balance between the sympathetic and parasympathetic nervous systems and the sensitivity of baroreflexes.

[0089] (3) Combining the slope features of the ascending branch, the steep features of the descending branch, and the relative height features of the diphthong notch in the pulse waveform morphology evolution trajectory after synchronous correlation and alignment processing, calculate the dynamic rate of change of pulse wave conduction velocity, and introduce this dynamic rate of change as a correction factor into the weighted calculation of HRV frequency domain index to generate a composite HRV feature vector containing vascular compliance compensation information. (4) Finally, the above time-domain indicators, frequency-domain indicators and composite HRV feature vectors are integrated to form HRV data corresponding to the evolution trajectory of heartbeat waveform morphology after synchronous correlation and alignment processing.

[0090] S150, the HRV evaluation network model is trained using the HRV data corresponding to the evolution trajectory of heartbeat waveform morphology features after synchronous correlation and alignment processing, the evolution trajectory of pulse waveform morphology features, and the evolution trajectory of heartbeat waveform morphology features after synchronous correlation and alignment processing.

[0091] In some embodiments, the above training process can be specifically described as follows: (1) Construct a dual-branch spatiotemporal fusion attention network model as the HRV evaluation network model. The dual-branch spatiotemporal fusion attention network model includes an ECG feature extraction branch, a PPG feature extraction branch, and a spatiotemporal fusion attention module. The ECG feature extraction branch is used to receive and process the evolution trajectory of the heartbeat waveform morphology features, the PPG feature extraction branch is used to receive and process the evolution trajectory of the pulse waveform morphology features, and the spatiotemporal fusion attention module is used to perform cross-modal correlation fusion of the ECG spatiotemporal features output by the ECG feature extraction branch and the PPG spatiotemporal features output by the PPG feature extraction branch.

[0092] Specifically, considering the significant temporal dependence and local morphological features of ECG signals, this branch adopts a cascaded architecture of a multi-layer one-dimensional convolutional neural network and a bidirectional gated recurrent unit network. The multi-layer one-dimensional convolutional neural network, by sliding convolutional kernels at different scales, can effectively capture local morphological features in the heartbeat waveform, such as the amplitude and width of the QRS wave. These local features constitute the basic units of the evolutionary trajectory of the heartbeat waveform morphological features. Subsequently, the bidirectional gated recurrent unit network, by introducing a gating mechanism and forward-backward bidirectional information flow, can perform contextual modeling of the local feature sequences extracted by the convolutional layers from two temporal directions, accurately capturing the dynamic evolution of the inter-heartbeat interval and the temporal dependence of heart rate changes, thereby outputting ECG spatiotemporal features rich in spatiotemporal information.

[0093] Correspondingly, the PPG feature extraction branch employs a completely symmetrical network structure. Since the PPG signal reflects changes in peripheral vascular blood flow volume, its waveform morphology features (such as the slope of the ascending branch, the steepness of the descending branch, and the relative height of the dicrotic notch) are closely related to the hemodynamic characteristics of cardiac ejection. By using a multi-layer one-dimensional convolutional neural network to finely characterize the local morphology of the pulse waveform, and then using a bidirectional gated recurrent unit network to perform temporal modeling of the continuous pulse interval, this branch can effectively extract the spatiotemporal features of PPG reflecting ventricular ejection strength, vascular status, and hemodynamics.

[0094] The spatiotemporal fusion attention module is the linchpin of the entire network model. Its core function is to perform deep cross-modal correlation and fusion of the ECG and PPG spatiotemporal features extracted from the two heterogeneous branches. This module integrates multi-head self-attention and cross-attention mechanisms: the multi-head self-attention mechanism first redistributes the weights of the temporal dependencies within the ECG and PPG features respectively, enhancing the representation of key information within each modality; the cross-attention mechanism achieves cross-modal information interaction and alignment by using the features of one modality as the query vector and the features of the other modality as the key-value pair. Through this bidirectional attention interaction, the network can adaptively learn the nonlinear coupling relationship between cardiac electrical activity and pulse mechanical fluctuations, explore the collaborative change patterns of the two in the spatiotemporal dimension, and ultimately generate a more information-rich and more representative fused spatiotemporal feature representation.

[0095] (2) Execute the first stage of joint pre-training process. Input the evolution trajectory of the heartbeat waveform morphology features after synchronous correlation and alignment processing into the ECG feature extraction branch, and input the evolution trajectory of the pulse waveform morphology features after synchronous correlation and alignment processing into the PPG feature extraction branch. Extract the spatiotemporal features of the heartbeat waveform morphology features evolution trajectory through the multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the ECG feature extraction branch to obtain ECG spatiotemporal features. Extract the spatiotemporal features of the pulse waveform morphology features evolution trajectory through the multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the PPG feature extraction branch to obtain PPG spatiotemporal features. Input the ECG spatiotemporal features and PPG spatiotemporal features into the spatiotemporal fusion attention module. The spatiotemporal fusion attention module uses multi-head self-attention and cross-attention mechanisms to perform feature interaction and weight allocation on ECG and PPG spatiotemporal features, generating a fused spatiotemporal feature representation. This fused spatiotemporal feature representation is then input into a fully connected regression layer, outputting predicted HRV data. A joint loss function is constructed based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronization association alignment processing, as well as the consistency loss between ECG and PPG spatiotemporal features. All network parameters of the dual-branch spatiotemporal fusion attention network model are iteratively updated using the backpropagation algorithm until the joint loss function converges to a preset threshold or reaches a preset number of training rounds, completing the first stage of joint pre-training.

[0096] Specifically, the first-stage joint pre-training process aims to utilize the synchronously aligned evolution trajectories of heartbeat waveform morphology and pulse waveform morphology to collaboratively optimize all parameters of the dual-branch spatiotemporal fusion attention network model through a multi-task learning strategy, thereby establishing an initial cross-modal correlation mapping between the heartbeat electrical signal and the pulse mechanical signal.

[0097] The process begins with data preparation and input. The evolution trajectory of the heartbeat waveform morphology features, after strict synchronization and alignment processing, is input into the ECG feature extraction branch. At the same time, the aligned pulse waveform morphology feature evolution trajectory is input into the PPG feature extraction branch. This ensures that the signals received by the two branches correspond precisely in the time dimension, reflecting the electromechanical coupling relationship under the same heartbeat cycle.

[0098] In the feature extraction stage, the multi-layer one-dimensional convolutional neural network in the ECG feature extraction branch abstracts the local morphological features of the heartbeat waveform layer by layer through stacked convolutional and pooling layers. It automatically learns physiologically meaningful waveform patterns from the initial raw sampling point sequence, and then feeds these local feature sequences into a bidirectional gated recurrent unit network. This recurrent network, with its gated structure and bidirectional information transmission capability, can capture the long-range dependence of the heartbeat interval from both forward and reverse directions, effectively modeling the dynamic trend and rhythmic fluctuations of heart rate, and ultimately outputting ECG spatiotemporal features that integrate local morphological details and global temporal evolution patterns. Simultaneously, the PPG feature extraction branch processes the evolution trajectory of the pulse waveform morphological features using the exact same forward computation logic, extracting PPG spatiotemporal features that characterize peripheral blood hemodynamics.

[0099] Subsequently, the spatiotemporal feature tensors output from these two branches are jointly input into the spatiotemporal fusion attention module. Within this module, a multi-head self-attention mechanism is first applied to independently enhance the ECG and PPG features, enabling each modality to recalibrate the importance weights of its features based on its internal temporal context. Next, a cross-attention mechanism is activated, constructing cross-modal attention connections so that ECG features can focus on key temporal segments within PPG features, and vice versa, thereby achieving deep interaction and information fusion of features in the spatiotemporal dimension. Through this combined effect of multi-head self-attention and cross-attention, the module generates a fused spatiotemporal feature representation that integrates complementary information from both modalities.

[0100] This fused feature is then passed to a fully connected regression layer, and after nonlinear mapping, the predicted HRV data is finally output. The prediction target can be a multi-dimensional HRV parameter vector including time-domain indicators (such as SDNN, RMSSD), frequency-domain indicators (such as LF, HF), and nonlinear indicators. The optimization objective of the model is guided by a joint loss function, which contains two core components: the first is the mean squared error loss between the predicted HRV data and the real HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing. This loss drives the model to learn an accurate mapping from the fused feature to the HRV indicator. The second is the consistency loss between the ECG spatiotemporal features and the PPG spatiotemporal features, which aims to constrain the representations of the two modalities in the feature space to be as close as possible, promote the network to learn modality-invariant feature representations, and enhance cross-modal semantic alignment. Through the backpropagation algorithm, the gradient of the joint loss function is used to iteratively update all network parameters in the dual-branch spatiotemporal fusion attention network model, including the feature extraction branch, fusion module, and regression layer.

[0101] This iterative process continues until the joint loss function converges to a preset threshold range or reaches the preset maximum number of training rounds, marking the successful completion of the first stage of joint pre-training. At this point, the network has initially grasped the complex relationship between ECG and PPG signals.

[0102] (3) Perform the second-stage independent fine-tuning process of the PPG feature extraction branch, freeze all network parameters of the ECG feature extraction branch or remove the ECG feature extraction branch from the dual-branch spatiotemporal fusion attention network model, input the pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment processing into the PPG feature extraction branch, extract spatiotemporal features of the pulse waveform morphology feature evolution trajectory through the multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the PPG feature extraction branch, obtain PPG spatiotemporal features, input the PPG spatiotemporal features into the spatiotemporal fusion attention module, and use the spatiotemporal fusion attention module to perform self-attention on the PPG spatiotemporal features. Feature enhancement involves generating PPG-enhanced spatiotemporal feature representations, which are then input into a fully connected regression layer to output predicted HRV data. A fine-tuning loss function is constructed based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology evolution trajectory after synchronous correlation alignment processing. The network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer are iteratively updated only through backpropagation until the fine-tuning loss function converges to a preset threshold or reaches a preset training epoch. This completes the second stage of independent fine-tuning of the PPG feature extraction branch, resulting in the final trained HRV evaluation network model.

[0103] Specifically, the second-stage PPG feature extraction branch independent fine-tuning process is a model lightweighting and specialization strategy designed based on the first-stage joint pre-training, which is only available for PPG signals in practical applications. Its core objective is to achieve single-modal reasoning capability that relies solely on PPG signals for high-precision HRV evaluation while maintaining cross-modal knowledge transfer performance.

[0104] This phase begins with adaptive adjustments to the model structure. Specifically, this involves freezing all network parameters of the ECG feature extraction branch to halt gradient updates, or, depending on deployment requirements, completely removing the ECG feature extraction branch from the dual-branch spatiotemporal fusion attention network model. The purpose of this freezing or removal is to preserve the feature extraction patterns learned in the first phase—guided by ECG signal supervision information and crucial for HRV prediction—while reducing computational redundancy and the number of model parameters, thus preparing for single-modal deployment.

[0105] For data input, the evolution trajectory of the pulse waveform morphology features after synchronous correlation and alignment processing is used as the sole input data and fed into the PPG feature extraction branch. This branch uses the network weights that have been preliminarily optimized in the first stage of pre-training. Its internal multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network perform forward calculation on the evolution trajectory of the input pulse waveform morphology features, extracting the local morphological features of the pulse waveform and its temporal evolution law layer by layer, and finally outputting the PPG spatiotemporal features.

[0106] Unlike the first stage, the PPG spatiotemporal features obtained in this stage no longer interact cross-modally with the ECG spatiotemporal features. Instead, they are directly input into the spatiotemporal fusion attention module. To adapt to single-modal input scenarios, the spatiotemporal fusion attention module's functionality is reconfigured: the cross-attention mechanism originally used for cross-modal interaction is disabled or bypassed, and the module only utilizes its internal multi-head self-attention mechanism to perform feature enhancement processing on the input PPG spatiotemporal features. This self-attention enhancement enables the model to focus on the temporal patterns and morphological feature points within the PPG signal that are most strongly correlated with HRV, such as emphasizing the enhancement of frequency band information related to respiratory sinus arrhythmia or waveform notch features related to baroreflex sensitivity, thereby generating a PPG enhanced spatiotemporal feature representation with higher information density.

[0107] Subsequently, this enhanced feature representation is passed to a fully connected regression layer, outputting the predicted HRV data through a nonlinear mapping. Model optimization is driven by a fine-tuning loss function, which directly calculates the mean squared error between the predicted HRV data and the actual HRV data corresponding to the heartbeat waveform morphological feature evolution trajectory after synchronous association alignment processing. Notably, during backpropagation, gradient updates are strictly limited to optimizing the network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer. The pre-trained ECG feature extraction branch parameters from the first stage (if not removed) are completely frozen and do not participate in this update. This parameter update strategy ensures that the fine-tuning process can fully utilize the prior knowledge gained from the ECG-PPG cross-modal learning in the first stage, while allowing PPG-related modules to be finely adjusted for the single-modal HRV prediction task to adapt to the morphological variability unique to PPG signals.

[0108] The fine-tuning process continues through iterative optimization until the fine-tuned loss function converges to a preset threshold or reaches a preset number of training rounds. At this point, the model has successfully transferred and solidified the bimodal knowledge from the first stage into the single-modal PPG processing path. The finally trained HRV evaluation network model has the ability to accurately and robustly evaluate HRV from PPG signals alone, meeting the requirements for low power consumption and high accuracy in practical applications.

[0109] S160: Real-time acquisition of PPG signal time series collected by wearable device of target object; real-time construction of pulse waveform morphological feature evolution trajectory based on real-time acquisition of PPG signal time series; input of the real-time constructed pulse waveform morphological feature evolution trajectory into the trained HRV evaluation network model; output of corresponding HRV data.

[0110] The real-time construction of the pulse waveform morphology feature evolution trajectory can be found in the above-mentioned content. Specifically, it involves: acquiring the PPG signal time series of the target object collected by the wearable device in real time; performing successive pulse waveform structure segmentation on the PPG signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units; and constructing the pulse waveform morphology feature evolution trajectory based on the rising slope feature, falling steepness feature, and relative height feature of the diatonic wave notch of each pulse waveform unit in the pulse waveform unit sequence.

[0111] In summary, the embodiments disclosed herein achieve at least the following technical effects: By constructing and synchronously linking the evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features, the accuracy of ECG signals in cardiac electrical activity timing and the richness of PPG signals in hemodynamic morphology are fully utilized, achieving complementary advantages and deep alignment of multimodal data. Using this high-precision synchronous trajectory data to train the HRV assessment network model, the model can learn the complex nonlinear mapping relationship from microscopic morphological changes of PPG waveforms to HRV indicators. Thus, with only low-power real-time acquired PPG signals as input, HRV data with measurement accuracy close to that of ECG signals can be output.

[0112] This not only completely eliminates the continuous reliance on high-power ECG circuits and significantly extends the battery life of wearable devices, meeting users' needs for all-weather, long-term health monitoring, but more importantly, by utilizing the alignment and complementarity effects of multimodal data, it greatly suppresses the interference of ambient light, motion artifacts, and individual differences on PPG signals, improving the reliability of assessment in scenarios with poor waveform morphology stability. This results in high accuracy of HRV analysis based on PPG signals, which can accurately reflect the true state of autonomic nervous system regulation, providing a technical solution for clinical disease early warning and daily health management that combines low cost, low power consumption, and high confidence.

[0113] It is worth noting that, in order to solve the "time misalignment" problem caused by inconsistent sampling frequencies and different transmission delays of different sensors, and especially to eliminate motion artifacts, the above method 100 further includes the following before performing waveform structure segmentation processing on the multi-mode physiological signal time series: The system acquires the IMU signal time series of the target object collected by a wearable device (specifically, an IMU sensor) within a preset monitoring period. The IMU signal time series, ECG signal time series, and PPG signal time series have different sampling frequencies, but share a unified time base timestamp, which is either a Unix timestamp or a tick count after device startup.

[0114] Based on a unified time reference timestamp, the IMU signal time series, ECG signal time series, and PPG signal time series are resampled or interpolated to place the signal time series with different sampling frequencies onto the same time axis. Subsequently, based on cross-correlation analysis algorithms (LMS algorithm or RLS algorithm), the correlation between the periodic motion peaks in the IMU signal time series and the noise peaks caused by motion in the ECG and PPG signal time series is calculated to automatically compensate for the transmission delay differences of each signal time series. Furthermore, by detecting the time difference between the R-wave cardiac electrical contraction start point in the ECG signal time series and the peak pulse arrival time point in the PPG signal time series, the motion state calculated from the IMU signal time series is combined to dynamically correct the delay fluctuations caused by changes in vascular elasticity in the PPG signal time series, and finally, the synchronization of the IMU signal time series, ECG signal time series, and PPG signal time series is completed.

[0115] The synchronized IMU signal time series is used as the reference noise signal input, and the synchronized ECG signal time series and PPG signal time series are used as the main input signals. An adaptive filtering algorithm is used to perform motion artifact filtering on the synchronized ECG signal time series and PPG signal time series in order to remove the motion noise components related to the IMU signal time series from the synchronized ECG signal time series and PPG signal time series.

[0116] Based on this, in order to improve the model accuracy, the above-mentioned training of the HRV evaluation network model using the HRV data corresponding to the synchronously correlated and aligned heartbeat waveform morphological feature evolution trajectory, pulse waveform morphological feature evolution trajectory, and heartbeat waveform morphological feature evolution trajectory can be expanded as follows: A three-branch spatiotemporal fusion attention network model is constructed as the HRV evaluation network model. The three-branch spatiotemporal fusion attention network model includes an IMU feature extraction branch, an ECG feature extraction branch, a PPG feature extraction branch, and a spatiotemporal fusion attention module. The IMU feature extraction branch is used to receive and process the time series of synchronized IMU signals. The ECG feature extraction branch is used to receive and process the evolution trajectory of heartbeat waveform morphology features after synchronization and alignment. The PPG feature extraction branch is used to receive and process the evolution trajectory of pulse waveform morphology features after synchronization and alignment. The spatiotemporal fusion attention module is used to perform cross-modal correlation fusion on the IMU spatiotemporal features output by the IMU feature extraction branch, the ECG spatiotemporal features output by the ECG feature extraction branch, and the PPG spatiotemporal features output by the PPG feature extraction branch.

[0117] The first phase of joint pre-training involves inputting the synchronized IMU signal time series into the IMU feature extraction branch, the synchronized and aligned heartbeat waveform morphological feature evolution trajectory into the ECG feature extraction branch, and the synchronized and aligned pulse waveform morphological feature evolution trajectory into the PPG feature extraction branch. Spatiotemporal features are extracted from the input data using multi-layer one-dimensional convolutional neural networks and bidirectional gated recurrent unit networks in each feature extraction branch, yielding IMU spatiotemporal features, ECG spatiotemporal features, and PPG spatiotemporal features, respectively. These features are then input into the spatiotemporal fusion attention module, utilizing spatiotemporal fusion... The multi-head self-attention mechanism and cross-attention mechanism in the attention module perform feature interaction and weight allocation on multimodal spatiotemporal features to generate a fused spatiotemporal feature representation. The fused spatiotemporal feature representation is input into a fully connected regression layer and outputs predicted HRV data. Based on the mean square error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing, as well as the consistency loss of spatiotemporal features of each branch, a joint loss function is constructed. The backpropagation algorithm is used to iteratively update all network parameters of the three-branch spatiotemporal fusion attention network model until the joint loss function converges to a preset threshold or reaches a preset training round, thus completing the first stage of joint pre-training.

[0118] The second-stage multi-branch independent fine-tuning process is executed. All network parameters of the ECG and IMU feature extraction branches are frozen, or these branches are removed from the three-branch spatiotemporal fusion attention network model. The pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment is input to the PPG feature extraction branch. The multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the PPG feature extraction branch are used to extract spatiotemporal features from the pulse waveform morphology feature evolution trajectory, obtaining PPG spatiotemporal features. These PPG spatiotemporal features are then input to the spatiotemporal fusion attention module, which is used to process the PPG spatiotemporal features. Self-attention feature enhancement is performed to generate PPG-enhanced spatiotemporal feature representations. These PPG-enhanced spatiotemporal feature representations are then input into a fully connected regression layer, which outputs predicted HRV data. A fine-tuning loss function is constructed based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing. The network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer are iteratively updated only through the backpropagation algorithm until the fine-tuning loss function converges to a preset threshold or reaches a preset number of training rounds. This completes the second stage of independent fine-tuning of the PPG feature extraction branch, resulting in the final trained HRV evaluation network model.

[0119] In view of this, the above embodiments construct a three-branch spatiotemporal fusion attention network model. Through a two-stage training strategy of joint pre-training and independent fine-tuning, high-precision HRV evaluation based on single-modal PPG signals is achieved. Specifically: In the first stage, the model simultaneously fuses IMU, ECG, and PPG multimodal data. Utilizing the multi-head self-attention and cross-attention mechanisms in the spatiotemporal fusion attention module, it deeply mines the correlation and complementarity of cross-modal spatiotemporal features. This allows the PPG feature extraction branch to fully learn the heart rhythm and body movement context information contained in the ECG and IMU signals during joint training, significantly improving its feature extraction capability. In the second stage, the ECG and IMU feature extraction branches are frozen or removed, and only the PPG feature extraction branch and the spatiotemporal fusion attention module are fine-tuned. While preserving the multimodal knowledge transfer effect, the single-modal inference performance is optimized, ensuring that the model can output accurate HRV data relying solely on PPG signals in practical applications.

[0120] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0121] The above is an introduction to the method embodiments. The following describes the solution described in this disclosure further through device embodiments.

[0122] Figure 2 A structural diagram of an HRV status assessment device based on PPG data provided in this disclosure is shown, as follows: Figure 2 As shown, the device 200 may include: The acquisition module 210 is used to acquire the time series of multi-mode physiological signals of the target object collected by the wearable device within a preset monitoring period; The segmentation module 220 is used to perform successive heartbeat waveform structure segmentation processing on the ECG signal time series in the multi-mode physiological signal time series to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units; and to perform successive pulse waveform structure segmentation processing on the PPG signal time series in the multi-mode physiological signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units. The construction module 230 is used to construct the evolution trajectory of the heartbeat waveform morphology features based on the QRS wave features of each heartbeat waveform unit in the heartbeat waveform unit sequence; and to construct the evolution trajectory of the pulse waveform morphology features based on the slope features of the ascending branch, the steepness features of the descending branch, and the relative height features of the diphthong notch of each pulse waveform unit in the pulse waveform unit sequence. Alignment module 240 is used to synchronously align the evolution trajectory of heartbeat waveform morphology features and pulse waveform morphology features on the time axis; and to calculate the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous alignment. Training module 250 is used to train the HRV evaluation network model using the HRV data corresponding to the evolution trajectory of heartbeat waveform morphology features after synchronous correlation and alignment processing, the evolution trajectory of pulse waveform morphology features, and the evolution trajectory of heartbeat waveform morphology features after synchronous correlation and alignment processing. The evaluation module 260 is used to acquire the PPG signal time series of the target object collected by the wearable device in real time, construct the pulse waveform morphology feature evolution trajectory in real time based on the acquired PPG signal time series, input the constructed pulse waveform morphology feature evolution trajectory into the trained HRV evaluation network model, and output the corresponding HRV data.

[0123] Understandable Figure 2 Each module / unit in the illustrated device 200 has the ability to implement Figure 1 The functions of each step in method 100 shown, and their corresponding technical effects, will not be elaborated here for the sake of brevity.

[0124] Figure 3A structural diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Electronic device 300 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 300 may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0125] like Figure 3 As shown, the electronic device 300 may include a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0126] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0127] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as method 100. For example, in some embodiments, method 100 may be implemented as a computer program product, including a computer program tangibly contained in a computer-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of method 100 described above may be performed. Alternatively, in other embodiments, the computing unit 301 may be configured to perform method 100 by any other suitable means (e.g., by means of firmware).

[0128] The various embodiments described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), payload programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0129] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0130] In the context of this disclosure, a computer-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0131] It should be noted that this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions. These computer instructions are used to cause a computer to execute method 100 and achieve the corresponding technical effects achieved by executing the method in the embodiments of this disclosure; for the sake of brevity, they will not be elaborated further here.

[0132] In addition, this disclosure also provides a computer program product including a computer program that implements method 100 when executed by a processor.

[0133] To provide interaction with a user, the embodiments described above can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0134] The embodiments described above can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0135] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0136] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0137] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for HRV state assessment based on PPG data, characterized in that, The method includes: Acquire the time series of multimodal physiological signals of the target object collected by wearable devices within a preset monitoring period; The ECG signal time series in the multi-mode physiological signal time series is subjected to successive heartbeat waveform structure segmentation processing to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units; and the PPG signal time series in the multi-mode physiological signal time series is subjected to successive pulse waveform structure segmentation processing to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units. The evolution trajectory of heartbeat waveform morphology features is constructed based on the QRS wave characteristics of each heartbeat waveform unit in the heartbeat waveform unit sequence; and the evolution trajectory of pulse waveform morphology features is constructed based on the rising slope characteristics, falling steepness characteristics, and diphtheria wave notch relative height characteristics of each pulse waveform unit in the pulse waveform unit sequence. The evolution trajectories of the heartbeat waveform morphology features and the pulse waveform morphology features are synchronously correlated and aligned on the time axis; and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous correlation and alignment are calculated. The HRV evaluation network model was trained using the evolution trajectories of heartbeat waveform morphology features and pulse waveform morphology features after synchronous correlation and alignment processing, as well as the HRV data corresponding to the evolution trajectories of heartbeat waveform morphology features after synchronous correlation and alignment processing. The time series of PPG signals collected by wearable devices for the target object are acquired in real time. The evolution trajectory of pulse waveform morphology features is constructed in real time based on the acquired PPG signal time series. The real-time constructed evolution trajectory of pulse waveform morphology features is input into the trained HRV evaluation network model, and the corresponding HRV data is output.

2. The method according to claim 1, characterized in that, The step of performing successive heartbeat waveform structure segmentation on the ECG signal time series in the multi-mode physiological signal time series to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units includes: The ECG signal time series is preprocessed by bandpass filtering to remove baseline drift and high-frequency electromyography interference, generating a denoised and clean ECG signal time series. The clean ECG signal time series is subjected to multi-scale decomposition processing based on wavelet transform to extract detail coefficients corresponding to the ventricular depolarization frequency band, and all local energy maxima are identified using an adaptive threshold detection algorithm to generate an initial set of candidate R-wave peak locations. The initial set of candidate R-wave peak locations is subjected to dual screening processing based on amplitude and polarity. Candidates with amplitudes lower than the preset minimum detectable threshold of the electrocardiogram signal and those with polarity opposite to the dominant heart rhythm are removed to obtain an effective R-wave peak location sequence. Calculate the time interval between two adjacent R-wave peak positions in the effective R-wave peak position sequence to obtain the RR interval sequence; The physiological consistency verification analysis is performed on the RR interval sequence. The R wave peak position corresponding to the interval value in the RR interval sequence whose relative change rate with the previous interval value exceeds the preset allowable range of cardiac chronotropicity is removed from the effective R wave peak position sequence. The signal segment near the removed position is subjected to secondary differential zero-crossing detection to correct the missed and / or false R wave peak positions, so as to obtain the corrected accurate R wave peak position sequence. Based on the time coordinates of each R-wave peak position in the corrected precise R-wave peak position sequence, a backward search is performed in the clean ECG signal time series to locate the nearest local minimum of the signal amplitude before each R-wave peak position as the P-wave start point position of the current heartbeat. Simultaneously, the search continues backward to locate the point where the signal amplitude returns to the baseline level and the second derivative approaches zero after each R-wave peak, which is taken as the T-wave termination point of the current heartbeat. The complete signal segment between each precise R-wave peak position and its corresponding P-wave start position and T-wave end position is extracted from the clean ECG signal time series to obtain a single heartbeat waveform unit containing a complete P-QRS-T wave group structure. Arrange all the extracted single heartbeat waveform units in chronological order according to the position of their R-wave peaks to generate a heartbeat waveform unit sequence composed of consecutive single heartbeat waveform units.

3. The method according to claim 1, characterized in that, The step of performing successive pulse waveform structure segmentation on the PPG signal time series in the multi-mode physiological signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units includes: Local maximum point detection processing is performed on the PPG signal time series to identify the sampling point positions in the PPG signal time series whose amplitude is greater than the amplitude of the sampling point to its left and the sampling point to its right, and to generate an initial set of candidate pulse wave peak positions. The initial pulse wave peak candidate point location set is subjected to amplitude filtering processing to remove the peak candidate point locations in the initial pulse wave peak candidate point location set whose amplitude is lower than the preset minimum amplitude threshold of the detectable PPG signal, thereby obtaining an effective pulse wave peak point location sequence. Calculate the time interval between two adjacent peak positions in the effective pulse wave peak position sequence to obtain a successive pulse interval value sequence. A physiological rationality analysis is performed on the successive pulse interval numerical sequence. The position of the next peak point corresponding to the interval value in the successive pulse interval numerical sequence whose difference from the previous interval value exceeds the preset physiological fluctuation range is removed from the effective pulse wave peak point position sequence to obtain the corrected pulse wave peak point position sequence. Based on the time coordinates of each peak position in the corrected pulse wave peak position sequence, the nearest local minimum point of signal amplitude before each peak position is located in the PPG signal time sequence as the waveform start point of the pulse, and the nearest local minimum point of signal amplitude after each peak position is located as the waveform end point of the pulse. The signal segment between each peak position and its corresponding waveform start and end positions is extracted from the PPG signal time series to obtain a complete single pulse waveform unit. All the captured single pulse waveform units are arranged in chronological order of their peak positions to generate a pulse waveform unit sequence composed of consecutive single pulse waveform units.

4. The method according to claim 1, characterized in that, The step of constructing the evolution trajectory of heartbeat waveform morphology features based on the QRS wave features of each heartbeat waveform unit in the heartbeat waveform unit sequence includes: The first sampling point in each heartbeat waveform unit where the QRS wave signal amplitude is lowest and begins to rise is identified as the starting point of that heartbeat waveform unit. The sampling point in each heartbeat waveform unit where the QRS wave signal amplitude is highest is identified as the peak point of that heartbeat waveform unit. The last sampling point in each respiratory waveform unit where the QRS wave signal amplitude is lowest and ends its decline is identified as the ending point of that heartbeat waveform unit. Calculate the time interval between rising points, the time interval between peak points, and the time interval between ending points for adjacent heartbeat waveform units; The rising point time interval, peak point time interval, and termination point time interval are combined and spliced ​​in the feature dimension to form a heartbeat waveform morphological feature evolution trajectory containing rising point time interval information, peak point time interval information, and termination point time interval information.

5. The method according to claim 1, characterized in that, The step of constructing the pulse waveform morphological feature evolution trajectory based on the rising slope feature, falling steepness feature, and relative height feature of the diphthoplast notch of each pulse waveform unit in the pulse waveform unit sequence includes: The first sampling point with the lowest signal amplitude in each pulse waveform unit is identified as the waveform start point of that pulse waveform unit; the sampling point with the highest signal amplitude in each pulse waveform unit is identified as the waveform peak point of that pulse waveform unit; and the last sampling point with the lowest signal amplitude in each pulse waveform unit is identified as the waveform end point of that pulse waveform unit. Based on the time coordinates and amplitudes of the waveform's starting point and the waveform's peak point, calculate the average amplitude change per unit time from the waveform's starting point to the waveform's peak point, and use the average amplitude change per unit time as the rising slope feature corresponding to the pulse waveform unit. Based on the time coordinates and amplitudes of the waveform peak and the waveform termination point, the average amplitude change per unit time from the waveform peak to the waveform termination point is calculated, and the average amplitude change per unit time is used as the steep descending branch characteristic of the pulse waveform unit. In the descending phase of each pulse waveform unit, the first local minimum of the signal amplitude is searched after the peak of the waveform, and this local minimum of the signal amplitude is marked as the diatonic wave notch of the pulse waveform unit. Extract the amplitude of the diphthalamic wave notch, calculate the absolute value of the difference between the amplitude of the diphthalamic wave notch and the amplitude of the waveform peak, use the absolute value of the difference as the absolute height value of the diphthalamic wave notch corresponding to the pulse waveform unit, and divide the absolute height value of the diphthalamic wave notch by the amplitude of the waveform peak to obtain the relative height feature of the diphthalamic wave notch corresponding to the pulse waveform unit. According to the chronological order of each pulse waveform unit in the pulse waveform unit sequence, the rising slope feature corresponding to each pulse waveform unit is arranged sequentially to generate a first feature sub-trajectory of the rising slope feature changing over time; according to the chronological order of each pulse waveform unit in the pulse waveform unit sequence, the descending steepness feature corresponding to each pulse waveform unit is arranged sequentially to generate a second feature sub-trajectory of the descending steepness feature changing over time; according to the chronological order of each pulse waveform unit in the pulse waveform unit sequence, the dicrotic notch relative height feature corresponding to each pulse waveform unit is arranged sequentially to generate a third feature sub-trajectory of the dicrotic notch relative height feature changing over time. The first feature sub-trajectory, the second feature sub-trajectory, and the third feature sub-trajectory are combined and spliced ​​in the feature dimension to form a waveform morphology feature evolution trajectory that includes the evolution information of the slope of the ascending branch, the evolution information of the steepness of the descending branch, and the evolution information of the relative height of the diphthong notch.

6. The method according to claim 1, characterized in that, The step of synchronizing and aligning the evolution trajectories of the heartbeat waveform morphology features and the pulse waveform morphology features on the time axis includes: Extract the time coordinates of the QRS wave peaks corresponding to each heartbeat waveform unit from the evolution trajectory of the heartbeat waveform morphology features to generate a sequence of heartbeat event occurrence times; extract the time coordinates of the waveform peaks corresponding to each pulse waveform unit from the evolution trajectory of the pulse waveform morphology features to generate a sequence of pulse event occurrence times. For each pulse event occurrence time point in the pulse event occurrence time point sequence, find the heartbeat event occurrence time point in the heartbeat event occurrence time point sequence that has the shortest time distance from the pulse event occurrence time point, and determine the heartbeat event occurrence time point as the coupled heartbeat event time point corresponding to the pulse event occurrence time point.

7. The method according to claim 1, characterized in that, The training of the HRV evaluation network model using the heartbeat waveform morphological feature evolution trajectory after synchronous correlation and alignment processing, the pulse waveform morphological feature evolution trajectory, and the HRV data corresponding to the heartbeat waveform morphological feature evolution trajectory after synchronous correlation and alignment processing includes: A dual-branch spatiotemporal fusion attention network model is constructed as the HRV evaluation network model. The dual-branch spatiotemporal fusion attention network model includes an ECG feature extraction branch, a PPG feature extraction branch, and a spatiotemporal fusion attention module. The ECG feature extraction branch is used to receive and process the evolution trajectory of heartbeat waveform morphological features, the PPG feature extraction branch is used to receive and process the evolution trajectory of pulse waveform morphological features, and the spatiotemporal fusion attention module is used to perform cross-modal correlation fusion of the ECG spatiotemporal features output by the ECG feature extraction branch and the PPG spatiotemporal features output by the PPG feature extraction branch. The first phase of joint pre-training is performed by inputting the evolution trajectory of the heartbeat waveform morphology features after synchronous correlation and alignment processing into the ECG feature extraction branch, and inputting the evolution trajectory of the pulse waveform morphology features after synchronous correlation and alignment processing into the PPG feature extraction branch. The ECG feature extraction branch uses a multi-layer one-dimensional convolutional neural network and a bidirectional gated recurrent unit network to extract spatiotemporal features from the heartbeat waveform morphology feature evolution trajectory, obtaining ECG spatiotemporal features. Similarly, the PPG feature extraction branch uses a multi-layer one-dimensional convolutional neural network and a bidirectional gated recurrent unit network to extract spatiotemporal features from the pulse waveform morphology feature evolution trajectory, obtaining PPG spatiotemporal features. The ECG spatiotemporal features and the PPG spatiotemporal features are then input into the spatiotemporal fusion attention module, utilizing the... The multi-head self-attention mechanism and cross-attention mechanism in the spatiotemporal fusion attention module perform feature interaction and weight allocation on the ECG spatiotemporal features and the PPG spatiotemporal features to generate a fused spatiotemporal feature representation. The fused spatiotemporal feature representation is input into a fully connected regression layer to output predicted HRV data. Based on the mean square error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing, as well as the consistency loss of the ECG spatiotemporal features and the PPG spatiotemporal features, a joint loss function is constructed. The backpropagation algorithm is used to iteratively update all network parameters of the dual-branch spatiotemporal fusion attention network model until the joint loss function converges to a preset threshold or reaches a preset training round, thus completing the first stage of joint pre-training. The second-stage PPG feature extraction branch independent fine-tuning process is performed. All network parameters of the ECG feature extraction branch are frozen, or the ECG feature extraction branch is removed from the dual-branch spatiotemporal fusion attention network model. The pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment processing is input to the PPG feature extraction branch. Spatiotemporal features of the pulse waveform morphology feature evolution trajectory are extracted using a multi-layer one-dimensional convolutional neural network and a bidirectional gated recurrent unit network in the PPG feature extraction branch, resulting in PPG spatiotemporal features. These PPG spatiotemporal features are then input to the spatiotemporal fusion attention module, which performs self-attention on them. Feature enhancement involves generating a PPG-enhanced spatiotemporal feature representation, which is then input into the fully connected regression layer to output predicted HRV data. A fine-tuning loss function is constructed based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing. The network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer are iteratively updated only through backpropagation until the fine-tuning loss function converges to a preset threshold or reaches a preset training epoch. This completes the second stage of independent fine-tuning of the PPG feature extraction branch, resulting in the final trained HRV evaluation network model.

8. The method according to claim 1, characterized in that, The method further includes: The IMU signal time series of the target object collected by the wearable device within a preset monitoring period is obtained; wherein the sampling frequencies of the IMU signal time series, the ECG signal time series and the PPG signal time series are different, but they have a timestamp with a unified time reference. Based on the timestamp of the unified time reference, the IMU signal time series, the ECG signal time series, and the PPG signal time series are resampled or interpolated to resample or interpolate the signal time series with different sampling frequencies onto the same time axis. Then, based on the cross-correlation analysis algorithm, the correlation between the periodic motion peaks in the IMU signal time series and the noise peaks caused by motion in the ECG signal time series and the PPG signal time series is calculated to automatically compensate for the transmission delay differences of each signal time series. Furthermore, by detecting the time difference between the R-wave cardiac electrical contraction start point of the ECG signal time series and the peak pulse arrival time point of the PPG signal time series, the motion state calculated by the IMU signal time series is combined to dynamically correct the delay fluctuations caused by changes in vascular elasticity in the PPG signal time series, and finally, the synchronization of the IMU signal time series, ECG signal time series, and PPG signal time series is completed. The synchronized IMU signal time series is used as the reference noise signal input, and the synchronized ECG signal time series and PPG signal time series are used as the main input signals. An adaptive filtering algorithm is used to perform motion artifact filtering on the synchronized ECG signal time series and PPG signal time series in order to remove the motion noise components related to the IMU signal time series from the synchronized ECG signal time series and PPG signal time series. The step of performing successive heartbeat waveform structure segmentation on the ECG signal time series in the multi-mode physiological signal time series to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units includes: The filtered ECG signal time series is subjected to successive heartbeat waveform structure segmentation to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units. The step of performing successive pulse waveform structure segmentation on the PPG signal time series in the multi-mode physiological signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units includes: The filtered PPG signal time series is subjected to successive pulse waveform structure segmentation to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units.

9. The method according to claim 8, characterized in that, The training of the HRV evaluation network model using the heartbeat waveform morphological feature evolution trajectory after synchronous correlation and alignment processing, the pulse waveform morphological feature evolution trajectory, and the HRV data corresponding to the heartbeat waveform morphological feature evolution trajectory after synchronous correlation and alignment processing includes: A three-branch spatiotemporal fusion attention network model is constructed as the HRV evaluation network model. The three-branch spatiotemporal fusion attention network model includes an IMU feature extraction branch, an ECG feature extraction branch, a PPG feature extraction branch, and a spatiotemporal fusion attention module. The IMU feature extraction branch is used to receive and process the synchronized IMU signal time series. The ECG feature extraction branch is used to receive and process the heartbeat waveform morphological feature evolution trajectory after synchronization and correlation alignment. The PPG feature extraction branch is used to receive and process the pulse waveform morphological feature evolution trajectory after synchronization and correlation alignment. The spatiotemporal fusion attention module is used to perform cross-modal correlation fusion on the IMU spatiotemporal features output by the IMU feature extraction branch, the ECG spatiotemporal features output by the ECG feature extraction branch, and the PPG spatiotemporal features output by the PPG feature extraction branch. The first phase of joint pre-training is performed by inputting the synchronized IMU signal time series into the IMU feature extraction branch, the synchronized and aligned heartbeat waveform morphological feature evolution trajectory into the ECG feature extraction branch, and the synchronized and aligned pulse waveform morphological feature evolution trajectory into the PPG feature extraction branch. Spatiotemporal features of the input data are extracted using multi-layer one-dimensional convolutional neural networks and bidirectional gated recurrent unit networks in each feature extraction branch, yielding IMU spatiotemporal features, ECG spatiotemporal features, and PPG spatiotemporal features, respectively. These features are then input into the spatiotemporal fusion attention module, utilizing the... The multi-head self-attention mechanism and cross-attention mechanism in the spatiotemporal fusion attention module perform feature interaction and weight allocation on multimodal spatiotemporal features to generate a fused spatiotemporal feature representation. The fused spatiotemporal feature representation is input into a fully connected regression layer to output predicted HRV data. Based on the mean square error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing, as well as the consistency loss of each branch spatiotemporal feature, a joint loss function is constructed. The backpropagation algorithm is used to iteratively update all network parameters of the three-branch spatiotemporal fusion attention network model until the joint loss function converges to a preset threshold or reaches a preset training round, thus completing the first stage of joint pre-training. The second-stage multi-branch independent fine-tuning process is executed. All network parameters of the ECG feature extraction branch and the IMU feature extraction branch are frozen, or the ECG feature extraction branch and the IMU feature extraction branch are removed from the three-branch spatiotemporal fusion attention network model. The pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment processing is input to the PPG feature extraction branch. The multi-layer one-dimensional convolutional neural network and bidirectional gated recurrent unit network in the PPG feature extraction branch are used to extract spatiotemporal features of the pulse waveform morphology feature evolution trajectory, obtaining PPG spatiotemporal features. The PPG spatiotemporal features are then input to the spatiotemporal fusion attention module, which is used to process the PPG spatiotemporal features. The empty features are enhanced with self-attention features to generate PPG-enhanced spatiotemporal feature representations. These PPG-enhanced spatiotemporal feature representations are then input into the fully connected regression layer, which outputs predicted HRV data. Based on the mean squared error loss between the predicted HRV data and the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous association alignment processing, a fine-tuning loss function is constructed. The network parameters of the PPG feature extraction branch, the spatiotemporal fusion attention module, and the fully connected regression layer are iteratively updated only through the backpropagation algorithm until the fine-tuning loss function converges to a preset threshold or reaches a preset number of training rounds. This completes the second stage of independent fine-tuning of the PPG feature extraction branch, resulting in the final trained HRV evaluation network model.

10. A HRV (HRV status assessment) device based on PPG data, characterized in that, The device includes: The acquisition module is used to acquire the time series of multimodal physiological signals of the target object collected by the wearable device within a preset monitoring period; The segmentation module is used to perform successive heartbeat waveform structure segmentation processing on the ECG signal time series in the multi-mode physiological signal time series to obtain a heartbeat waveform unit sequence composed of continuous single heartbeat waveform units; and to perform successive pulse waveform structure segmentation processing on the PPG signal time series in the multi-mode physiological signal time series to obtain a pulse waveform unit sequence composed of continuous single pulse waveform units. The construction module is used to construct the evolution trajectory of heartbeat waveform morphology features based on the QRS wave features of each heartbeat waveform unit in the heartbeat waveform unit sequence; and to construct the evolution trajectory of pulse waveform morphology features based on the rising slope features, falling steepness features, and diphtheria wave notch relative height features of each pulse waveform unit in the pulse waveform unit sequence. The alignment module is used to synchronously align the evolution trajectory of the heartbeat waveform morphology features and the evolution trajectory of the pulse waveform morphology features on the time axis; and to calculate the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory after synchronous alignment. The training module is used to train the HRV evaluation network model using the HRV data corresponding to the heartbeat waveform morphology feature evolution trajectory and pulse waveform morphology feature evolution trajectory after synchronous correlation and alignment processing. The evaluation module is used to acquire the PPG signal time series of the target object collected by the wearable device in real time, construct the pulse waveform morphology feature evolution trajectory in real time based on the acquired PPG signal time series, input the constructed pulse waveform morphology feature evolution trajectory into the trained HRV evaluation network model, and output the corresponding HRV data.

Images